Основные понятия
Intent classification is a crucial component of dialogue systems, but faces significant challenges in adapting to new domains. This review analyzes contemporary datasets, methods, and limitations to enable more effective and adaptable intent classification.
Аннотация
This paper provides a comprehensive review of intent classification systems for dialogue agents. The authors first analyze the datasets used to train intent classifiers, covering aspects like data type, multilingualism, and domain coverage. They categorize the contemporary methods for intent classification into three main approaches: fine-tuning of pre-trained language models (PLMs), prompting of PLMs, and few-shot/zero-shot learning.
The authors then discuss why intent classification is a difficult task, highlighting challenges like the multimodal nature of human communication, the need for customizability, the lack of reasoning ability in PLMs, the diversity of natural language, the similarity of intents, the lack of training data, imbalanced training data, and out-of-vocabulary issues.
Based on these challenges, the authors identify several open issues that deserve more attention from NLP researchers to improve the adaptability of intent classification systems. These include the need for multimodal input data, limitations of existing datasets, the resource-intensive nature of LM fine-tuning, the challenges of GPT-prompting for semantically-close intents, and the language dependence of current systems.
The authors conclude by outlining future directions to address these limitations, such as creating multimodal and multilingual datasets with diverse domains, exploring conversational pretraining objectives and adapter-based approaches, and leveraging contrastive learning for few-shot classification of intents.
Статистика
"Dialogue agents continue to draw attention of NLP researchers, leading to the development of several methods, datasets, and objectives needed to train agents to classify user-intent while performing a task."
"To achieve effective dialogue agents, the implementation of intent classification involves deploying NLU systems to identify the intent of the user."
"Dialogue agents should adapt easily from one domain to another, for them to be more effective."
Цитаты
"To achieve such systems, researchers have developed a broad range of techniques, objectives, and datasets for intent classification."
"Despite the progress made to develop intent classification systems (ICS), a systematic review of the progress from a technical perspective is yet to be conducted."
"Herein, intent classification predicts the intent label of the query."